Spaces:

Slep

/

CondViT-LRVSF-Demo

like 5

Running

Hi, thanks for your interest and the heads up about the demo !
There was a build error, it should be up an running again now :)

Hi guys ... sure i'm interested. Good paper , in any case I'll try to run it , as you indicated in the project page:
I have 2 small questions about it : a) do u have a checkpoint of the training? because for me it's a bit difficult to retrain, I use an academic server very busy b) a silly question about the test: why to embed and after evaluate the similarity? " 2 separate steps. There is some reasons about architecture?
Thanks in advance. Paolo

a) Yes, there is a checkpoint of a Categorical CondViT-B/16 in the demo files (artifacts/cat_condvit_b16.pth). If you look at the demo code you should easily be able to use it.

b) The main reason is that computing the embeddings can take a while when you have millions of images, and I don’t want to do it multiple times. So I precompute the embeddings once for the whole test set (which takes a while), and store them. Then I can load parts of the test set, compute multiple metrics, use it as an index for a demo, or whatever, and it takes only a few seconds each time with optimized libraries like FAISS.

I hope it answers your questions.

Hi
a) Good. I didn't find that (i have read quickly. sorry)
b) yes, obvious, all embedded data once, it has been a stupid question, i must go dive in the code :). Yes thank for fully info. i'll keep you updated. Good day P

Hi Simon...Sorry, another question: the checkpoint used in the demo gradio is "./artifacts/cat_condvit_b16.pth", but I have found also B32_Params used for training in the python class of the main repository (I didn't find the related checkpoint, is it non public?). I don't know if the B32 model has more precision (the img have better quality), large training set or both?
Thks in advance. Paolo

Hi,
I didn't release the checkpoints for B/32 models, as I used this size mostly for ablations studies. Both B/32 and B/16 work on 224x224 images, but B/32 works on a coarser representation (larger patches) so it performs worse (cf Table 1 in the paper). Otherwise they have the same number of parameters and are trained on the same dataset.

Thanks for details Simon, I look the large patches. Very useful your answers, i hope give you something of useful. If you want ask me, I'm experience in DL and ML, more in NLP . good day P